基于核相关滤波器的目标跟踪方法研究
发布时间：2018-08-21 08:37
【摘要】：目标跟踪在现代机器视觉起着重要作用。最近基于核相关滤波的跟踪达到很好的效果,但仍有改进的必要,本文对其进行了全面分析,提出两个改进方案。为了解决遮挡引起的跟踪丢失问题,建立一个目标检测模块。首先通过对当前一些检测算法的分析,确定检测方案。通过提取目标模板与当前帧的SURF特征点进行匹配,运用改进的RANSAC算法过滤匹配对,计算变换矩阵,来定位当前帧的目标位置,以达到检测目的。同时建立模板集合来增加鲁棒性。建立了判断机制,通过在当前帧训练跟踪器,进行反向跟踪,然后比较结果判断是否启动检测,同时修改了模板更新方法。针对核滤波跟踪无法适应目标尺度变化的问题,通过引入目标候选框算法来产生尺度不同的方框。通过结构化随机森林引出提取目标候选框的edge box算法,同时修改了算法以适应需要。用原算法进行粗跟踪,在结果位置处提取区域,在该区域运行edge box以产生目标候选方框,将一些评分高的方框提取出来,结合原跟踪框进行筛选后,再变换回初始大小,然后代入KCF进行评估,综合原结果得出当前帧最合适的跟踪框。同时融合了多种特征以进一步提高整体跟踪效果。本文同时论述了跟踪评估方式,评估两种改进方案时从OTB数据集中分别选取29个遮挡属性的视频与28个尺度属性视频,分别通过定性与定量实验与原算法以及Stuck与TLD进行了对比。实验结果表明,本算法在成功率图与精确度图排名上均优于原KCF,TLD,struck算法。与原方法相比,改进后的方法能更好地适用于有尺度变化与遮挡的跟踪,能够广泛应用于目标跟踪领域。
[Abstract]:Target tracking plays an important role in modern machine vision. Recently, the tracking based on kernel correlation filter has achieved good results, but there is still a need for improvement. This paper makes a comprehensive analysis of it and puts forward two improved schemes. In order to solve the problem of tracking loss caused by occlusion, a target detection module is established. First of all, through the analysis of some current detection algorithms, the detection scheme is determined. By extracting the target template to match the SURF feature points of the current frame, the improved RANSAC algorithm is used to filter the matching pairs, and the transform matrix is calculated to locate the target position of the current frame so as to achieve the purpose of detection. At the same time, the template set is established to increase robustness. A judgment mechanism is established, and the method of template updating is modified by training the tracker in the current frame for reverse tracking, then comparing the results to judge whether the detection is initiated or not. In order to solve the problem that the kernel filter tracking can not adapt to the change of target scale, the target candidate algorithm is introduced to generate different scale boxes. The edge box algorithm for extracting target candidate is obtained by structured random forest, and the algorithm is modified to meet the needs. The original algorithm is used for rough tracking, the region is extracted at the result location, the edge box is run in this region to produce the target candidate box, and some highly graded boxes are extracted. After the original tracking box is filtered, the initial size is then transformed back to the initial size. Then the KCF is used to evaluate the current frame and the most suitable tracking box is obtained by synthesizing the original results. At the same time, a variety of features are fused to further improve the overall tracking effect. At the same time, this paper discusses the method of tracking and evaluation. When evaluating the two improved schemes, the video of 29 occlusion attributes and 28 scale attribute videos are selected from the OTB dataset, respectively. The qualitative and quantitative experiments are compared with the original algorithm, and the Stuck and TLD are compared. The experimental results show that the proposed algorithm is superior to the original KCFC TLDLD-struck algorithm in the ranking of the success rate diagram and the accuracy chart. Compared with the original method, the improved method is more suitable for tracking with scale variation and occlusion, and can be widely used in the field of target tracking.
【学位授予单位】：西南科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
[Abstract]:Target tracking plays an important role in modern machine vision. Recently, the tracking based on kernel correlation filter has achieved good results, but there is still a need for improvement. This paper makes a comprehensive analysis of it and puts forward two improved schemes. In order to solve the problem of tracking loss caused by occlusion, a target detection module is established. First of all, through the analysis of some current detection algorithms, the detection scheme is determined. By extracting the target template to match the SURF feature points of the current frame, the improved RANSAC algorithm is used to filter the matching pairs, and the transform matrix is calculated to locate the target position of the current frame so as to achieve the purpose of detection. At the same time, the template set is established to increase robustness. A judgment mechanism is established, and the method of template updating is modified by training the tracker in the current frame for reverse tracking, then comparing the results to judge whether the detection is initiated or not. In order to solve the problem that the kernel filter tracking can not adapt to the change of target scale, the target candidate algorithm is introduced to generate different scale boxes. The edge box algorithm for extracting target candidate is obtained by structured random forest, and the algorithm is modified to meet the needs. The original algorithm is used for rough tracking, the region is extracted at the result location, the edge box is run in this region to produce the target candidate box, and some highly graded boxes are extracted. After the original tracking box is filtered, the initial size is then transformed back to the initial size. Then the KCF is used to evaluate the current frame and the most suitable tracking box is obtained by synthesizing the original results. At the same time, a variety of features are fused to further improve the overall tracking effect. At the same time, this paper discusses the method of tracking and evaluation. When evaluating the two improved schemes, the video of 29 occlusion attributes and 28 scale attribute videos are selected from the OTB dataset, respectively. The qualitative and quantitative experiments are compared with the original algorithm, and the Stuck and TLD are compared. The experimental results show that the proposed algorithm is superior to the original KCFC TLDLD-struck algorithm in the ranking of the success rate diagram and the accuracy chart. Compared with the original method, the improved method is more suitable for tracking with scale variation and occlusion, and can be widely used in the field of target tracking.
【学位授予单位】：西南科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
【相似文献】
相关期刊论文 前10条
1 马进;田涛;;基于联合变换相关的目标跟踪方法[J];应用光学;2012年05期
2 尤小泉;彭映杰;;一种基于指令预测的目标跟踪方法[J];电视技术;2010年02期
3 邵文坤;黄爱民;韦庆;;目标跟踪方法综述[J];影像技术;2006年01期
4 梁德群;阮文;;基于模型的线性组合目标跟踪方法[J];模式识别与人工智能;1995年04期
5 王思，
本文编号：2195162
本文链接：https://www.wllwen.com/shoufeilunwen/xixikjs/2195162.html
